Built where the rivers decide the year.
ModelEarth is one engineer in Odisha, building the warning system he wishes had existed. Everything on this site was written, deployed and audited by the same person.
The technology already existed in 1999.
A super cyclone crossed the Odisha coast in October 1999 and took thousands of lives before a single warning reached the last village. The models existed. The satellites were overhead. What was missing was the final step.
Twenty-seven years later the rivers still rise on a schedule everyone knows, and the last mile is still the hard part. Not the physics - the distribution. That last mile ends at a district officer at three in the morning, with half the data and all of the responsibility, deciding whether to wake a town. Everything we build is aimed at that one moment.
“I grew up where the rivers decide the year. If we can buy one family one more day, the whole thing was worth building.”
Swayam Debata
Founder
Software engineer, born and raised in Odisha. Built the ingestion, the rule engine, the ML pipeline, the API, the dashboard, the alerting and the deploy.
Anil Kumar Moharana
Co-founder
No outside capital, raising a 75 lakh rupee pre-seed. Hiring next: one ML and data engineer, and one partnerships lead.
Where things actually stand
- Founders
- 2
- Outside capital raised
- None yet
- Paying customers
- 0
- Signed MoUs
- 0
- Cities alerting in production
- 5
- Trust gate
- Open
If you are an investor or a district and you were expecting a longer list, that is the point. The engine is real and the traction is early, and we would rather you learn both from us than from diligence.
Key-person risk is real here and we are not pretending otherwise. It is also why the discipline is written into the code as recorded decisions rather than kept in someone’s head: shadow cannot alert because a test fails if it ever could.
Four rules we do not bend.
Written into the codebase as recorded decisions, not aspirations on a wall. Breaking one requires a new entry in the decision log, with a reason.
D001
Five cities until the trust gate passes
Expanding the alert map is the easiest way to look bigger and the fastest way to page a district officer about a location we have never validated. New geography goes to shadow first, every time.
D002 · D005
Offline metrics are not permission to go live
The ML model has good numbers on held-out data. It has never driven an alert and will not until it survives a monsoon in shadow. A good score on history is a hypothesis.
D019
Shadow and live never share a screen
Flood Ops and Heat Ops are separate surfaces. A shadow location cannot alert by construction, and a test fails loudly if a refactor ever merges the two location lists.
D023 · D025
Publish the audit that costs you the number
We retired our own 99.3% headline the week we found it was mislabelled, and put the replacement - and the precision problem it exposed - on the public site. That is the whole culture in one move.
Two people, and they are not the same person.
A warning that arrives in time is the difference between a story and a statistic. But the warning has to reach two very different people, and pretending they need the same interface is how early-warning products fail.
The officer on watch
Needs a defensible decision at 3 a.m. and a paper trail at 9 a.m. Gets the War Room, Evidence Mode, and a replay they can run themselves.
The family at the river's edge
Needs one sentence, in their language, on a phone that may be the only one in the house. Gets KrishiOS: voice-first, six languages, shareable on WhatsApp.
Every score, every alert, every replay points back to the home at the river’s edge and the family that has nowhere higher to go.
A pilot is one district, one monsoon.
We score your basin against its own thirty-year baseline and run it alongside whatever you use today, so the comparison is yours to keep either way. Tell us the district and we will say plainly whether the record there supports a warning yet.
Or reach us directly